Ep 63 - Human In The Loop: Designing The Boundary Between Machines And Humans

Send us Fan Mail The moment an AI agent can issue refunds or change accounts, the conversation shifts from capability to responsibility. We dig into how to design trust between people and machines by choosing the right oversight model for the job: human in the loop for high-stakes decisions and human on the loop for fast, high-volume work. Along the way, we unpack concrete playbooks for customer service leaders and operators who need speed without sacrificing judgment. We start by drawing a ...

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Jason McGuinthe?

  Welcome back to Inspire AI, the podcast where we bring you closer to the world of artificial intelligence and how it's reshaping work, leadership, and decision-making. A real podcast for builders, operators, leaders who want more than hype. Each week, we translate what's happening in AI into practical mental models, real-world patterns, and next steps that you can actually use. Over the last couple of episodes, we've been exploring a powerful shift in how we build systems, AI agents. We talked about what agents are, how they reason, how they plan, and how we, as practitioners,

• 0:45Jason McGuinthe?

  work with them. We are all practitioners of AI now. We must learn how to delegate tasks, set goals, design guardrails, and integrate them into real workflows. And if you've been listening closely, there's a question sitting just beneath the surface. At what point do we stop letting AI act on its own? And when do humans step back in? Because once an AI agent isn't just answering questions, but issuing refunds, changing accounts, updating profiles, approving actions.

• 1:21Jason McGuinthe?

  Autonomy stops being exciting and starts being risky. So today's episode is the natural next step in this journey. If the first episodes were about building and collaborating with AI agents, then this one, it's about trust, accountability, and control. Today, we're diving into human in the loop and its counterpart, human on the loop, especially in customer service, where decisions are fast, emotional, and sometimes expensive. Imagine this scenario: An AI customer service agent approves a $50,000 refund to a fraudulent account automatically.

• 2:03Jason McGuinthe?

  No human approved, no review, just confidence, and a very angry CFO the next morning. It's not a failure of AI capability. That is a failure of system design. So today, we're going to break down what human in the loop and human on the loop really mean, how to choose between them based on risk, the architectures and triggers that make them work, some real world examples of teams doing it well,

• 2:33Jason McGuinthe?

  and what's coming next as regulation and best practices mature. So let's dig in. We know that AI agents are evolving and they're becoming part of our lives. We know a bit about what they do, how they operate, a little bit under the hood. We've already explored how we're going to think about working with them. But I think we're going to hit a wall. I really do. You see, people are going to start building agents and recognize that it reasons well, it executes tasks, and it can handle real customer interactions.

• 3:07Jason McGuinthe?

  Then what? The question becomes, who's accountable when it makes a decision? I don't care what industry you're in. I'm pretty sure that that question's going to be important to everyone, regulated or unregulated industries. But that's where a human in the loop and human on the loop come in. What is human in the loop? Well, it means that a human actively participates in the decision-making process. The AI might analyze the situation, draft a response, recommend an action, but nothing final happens without human approval. Think of this as co-piloting. Co-piloting, like in the original sense, there's a human pilot behind the wheel, but the AI flies and the human authorizes the key maneuvers.

• 3:50Jason McGuinthe?

  This is ideal when the cost of error is high. Decisions are irreversible. Context or empathy matters. Refunding approvals, closing accounts, handling sensitive complaints. This is the human in the loop territory, all right? Now, human on the loop, that's a bit different. The AI can act autonomously most of the time, while a human monitors performance, reviews logs and metrics, intervenes only when something looks wrong. It's supervisory control. Think back to my previous episodes about AI augmentation. There's a series of steps of maturity from AI having least privilege to AI having most privilege.

• 4:37Jason McGuinthe?

  Human on the Loop takes it to one of the highest forms of maturity. Back to the pilot analogy. Think of air traffic control. Planes fly themselves, but humans are watching every move. You ever heard of a drone? Hmm. Human on the Loop works best when volume is high. Doesn't make economic sense to put that many people in the loop. Errors are low risk or reversible. And speed matters. So, consider the previous episodes where we talked about delegating work to agents. Human in the Loop and Human on the Loop are how you bound that delegation, creating a contract.

• 5:21Jason McGuinthe?

  You must only stay within these guardrails. With Human in the Loop and Human on the Loop, you can now answer, how much autonomy is safe? When should humans intervene? How do we scale without losing control? So, beyond the fact that it's a kind of a philosophical idea, there's an architectural decision here. One of the most effective patterns is confidence-based escalation, where in moments of high confidence, the AI acts automatically. In moments of low confidence, human review is required. For example, a billing frequently asks question at 98% confidence. Hey, why not auto reply?

• 6:03Jason McGuinthe?

  But a refund request where there's emotional language and ambiguity involved, that should require some human review. This sort of system creates dynamic autonomy. It's not an all or nothing automation. Here's a mental model on oversight. For synchronous, Human in the Loop, where the AI will send something out and wait to get something back, the AI will pause, wait for approval before acting. It's slower, but it's safest. For asynchronous, Human on the Loop, AI acts immediately, but humans review outcomes later. It's faster, but it requires strong monitoring.

• 6:42Jason McGuinthe?

  And back to being dynamic, most mature systems would use both, depending on risk. Advanced teams would separate planning, where it's human approved, from execution, AI-driven. Humans approve the strategy.

• 6:59Jason McGuinthe?

  AI handles the steps. This is exactly how experienced teams scale without losing judgment. I think we should ground this in some examples. A global fashion retailer deploys an AI chatbot that resolved over 40% of inquiries automatically, but escalated emotionally charged or uncertain cases to humans. The result: faster responses, higher customer satisfaction, and humans focused on what mattered most. How about an airline using AI to suggest responses to human agents? Humans corrected it, the AI learned. Over time, simple questions became fully automated with supervisors watching in the background.

• 7:43Jason McGuinthe?

  They didn't just jump to autonomy, they learned it and they earned it. Here's a good one: insurance claims. AI can process most claims instantly. Humans reviewed edge cases and exceptions. Customers can get faster payouts. Regulators got auditability. Trust stays intact. So here's what I would suggest as you work through this. Beyond just mapping your tasks, map your risk. Automate low-risk actions first. Design clear escalation rules. What are your confidence, sentiment, and value thresholds? Give human context and explanations. Never make them guess why the AI flagged something.

• 8:27Jason McGuinthe?

  Turn human corrections into training data. Every override is a learning opportunity. Think about that. Watch for automation bias. If humans never disagree with AI, something might be wrong. Build governance in from day one. Logs, audits, override authority, kill switches. Train humans as AI supervisors, not button pushers. Think about that. How many times have you just accepted on first pass what the AI outputs to you? Giving up your control leads to giving up your critical thinking skills, or vice versa. I think that wherever this is heading, it's pretty exciting.

• 9:11Jason McGuinthe?

  AI systems that ask for help when uncertain, continuous learning from human feedback, supervisors managing fleets of agents? Yikes. Better confidence estimation and risk modeling, regulation-ready systems by design. We should be moving our mental model from AI can do this to how do we design trust between humans and AI? That is a great question. So if we zoom out, AI agents are about capability. Working with AI agents is about collaboration. Human in the loop design is about responsibility. It's humans guiding, teaching, and supervising machines at scale.

• 9:56Jason McGuinthe?

  Best AI systems don't remove humans from the equation. They elevate them. They let AI handle speed and repetition, while humans provide judgment, empathy, and accountability. As you build your own systems, remember autonomy isn't binary. Trust is earned. In the loop, that's where intelligence becomes wisdom. If you like this episode, please share it with a colleague, friend, family member. Thanks for joining us on Inspire AI. Remember to stay curious, design responsibly, and build AI that knows when to ask for human judgment.